fix(derivations): one bin per distinct value when more bins are requested than values exist

[cnicholas]

Summary

• What: when at least as many bins are requested as there are distinct finite values, fit the edges at the midpoints between neighbouring values, one bin per value, instead of capping the count and re-taking quantiles.
• Why: the 0.3.1 cap collapsed heavily tied columns. A column of 3,000 ones and 1,000 twos with n=4 fitted one bin, where 0.3.0 fitted two, because the halved quantile positions both landed on 1. Caught by the app's own preview test during the pin bump. Requesting far more bins than values (the case the cap was for) now also gives one bin per value instead of mostly-empty interpolated bins.
• Scope: _fit_edges in derivations.py; two existing tests re-pinned to the fewer-bins-than-values case they were written to prove; robustness tests extended; CHANGELOG.

Contract / invariants

• Defaults unchanged — requesting fewer bins than distinct values takes the original quantile / equal-width path with byte-identical edges (pinned). Every finite value still lands in a bin (fuzz and bin-extremes suites). validation/e2e_bishop_report.py exits 0.
• Residuals unaffected.
• Row/index alignment preserved.
• Output schema compatible.
• Pinned error strings unchanged — message wording for the collision case is unchanged; the one-per-value case has its own message.

Behaviour changes

• n >= distinct count: edges are [v1, midpoints…, vk] and the message reads requested n bins, ties produced k (one per distinct value). Labels change from interpolated quantile edges such as [1, 1.25) to midpoint edges such as [1, 1.5); bin membership is one value per bin either way.

Methodology

• No methodology change.

Tests

• test_heavily_tied_column_keeps_one_bin_per_distinct_value — the regression, 3,000 ones and 1,000 twos.

• test_more_bins_than_distinct_values_gives_one_bin_per_value — edges and counts for both methods.

• test_fewer_bins_than_distinct_values_is_untouched — the original path pinned.

• test_tie_drop_uses_fitted_count_labels_and_message and test_validate_label_count_against_fitted_not_requested — now request 4 bins on 5 distinct values so the quantile-collision case they document still occurs (5 on 5 now correctly fits 5).

• pytest tests/ — 2431 passed, 10 skipped

• ruff check . — clean

• Golden masters untouched

🤖 Generated with Claude Code

https://claude.ai/code/session_01BKphCi1LC1hjGYteWEK9JV